OpenAI Unveils New 'Reasoning' AI Models
OpenAI, the creator of ChatGPT, unveiled a new series of artificial intelligence (AI) models on Thursday aimed at allocating more time for processing. The objective is to enhance the accuracy and utility of responses generated by AI chatbots.
These improved models have undergone training to enhance their problem-solving skills, including experimenting with different strategies and identifying errors before finalizing their responses.
OpenAI CEO Sam Altman celebrated the release, saying these models represent "a new paradigm: AI that can do general-purpose complex reasoning." Yet, he also offered a note of caution, stating, "is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it."
The development aims to mitigate the issue of "hallucinations" in AI chatbots, where the systems generate convincing but incorrect answers—a problem that has tempered enthusiasm for ChatGPT-like AI among commercial users.
According to OpenAI researcher Jerry Tworek, who spoke to The Verge, there is a notable reduction in hallucinations with the new model, although he clarified, "we can't say we solved hallucinations."
In testing phases, the models have shown promising results, performing on par with PhD students in challenging physics, chemistry, and biology problems. They also demonstrated strong proficiency in mathematics and coding, successfully answering 83 percent of questions on an International Mathematics Olympiad qualifying exam, a significant leap from the 13 percent success rate of the previous model, GPT-4o.
OpenAI has suggested potential applications for these advanced reasoning abilities, ranging from assisting healthcare researchers in annotating cell sequencing data to aiding physicists in crafting complex formulas and supporting computer developers in creating and implementing multistep designs.
Debra A Smith for TROIB News